A cross-learning approach for cold-start forecasting of residential photovoltaic generation

This paper addresses the problem of forecasting, over a daily horizon, quarter hourly profiles of residential photovoltaic (PV) power production for sites with no historical data available. Typically, such forecasts are required for improving the local operation of low-voltage systems, where observability is still a practical challenge. In this context, we develop a cross-learning forecasting approach to predict unobserved PV sites, which exploits common patterns learned from neighboring monitored PV production profiles. Concretely, the proposed approach fits a single, generic forecasting function across the entire panel of monitored PV time series based only on series-specific features - i.e., the peak power installed, geographical position, orientation and inclination - and local numerical weather predictions. This allows to enlarge the dataset for training more complex data-driven techniques, while ensuring scalability for predicting each PV site. The proposed approach is evaluated using a k-nearest neighbors algorithm, different variants of neural networks and gradient boosted trees on five new residential PV sites. Outcomes highlight the ability of the cross-learning forecasting models to better generalize on new PV sites in comparison with a clear sky-based physical approach, without needing any adjustment of the models.